AI Transparency

The short version

Everything dub44 produces is synthetic. It is generated by a machine learning model from a recording of a voice — it is not a recording of anyone actually saying those words. Every file we hand you carries an inaudible watermark identifying it as AI-generated, and you cannot switch that off.

What the model is

We run Chatterbox, an open-source text-to-speech model from Resemble AI, on our own GPUs in the EU. It does zero-shot voice cloning: it takes a short reference recording, extracts what makes that voice sound like itself, and uses it to speak new text. No model is trained on your voice, and your recording is never used to improve anything we ship.

Watermarking

Every clip from every engine passes through Resemble’s Perth watermarker before it reaches you. The mark is inaudible, survives normal encoding and sharing, and lets the audio be identified as machine-generated after the fact. It is applied by the model itself, so there is no path through our system that produces an unmarked file.

Trial audio carries a second, audible spoken tag at the start. That one is a product limit rather than a compliance measure, and it goes away on a paid plan. The inaudible mark stays on every plan, forever.

Your obligations when you publish

Under Article 50 of the EU AI Act, whoever publishes synthetic audio has to disclose that it is artificially generated. That duty sits with you, not with us — we cannot know where your audio ends up. If you publish dub44 output in a context where a listener could reasonably mistake it for a real recording of a real person, say that it is AI-generated.

Fiction, satire and art have lighter obligations, but the disclosure still has to be there in some appropriate form. Using generated audio to make someone appear to say something they did not say is prohibited outright by our Acceptable Use Policy, regardless of what the law requires.

Consent

Clone your own voice, or one whose owner has explicitly agreed. You confirm this each time you create a voice, and we keep that confirmation. A voice is personal data and in many cases biometric data — cloning someone without their agreement is both a GDPR problem and, under our terms, grounds for termination.

What the model gets wrong

Honest limitations, so you are not surprised:

• It clones timbre far better than accent or rhythm. A clone of a non-native English speaker often keeps the voice but loses the accent.
• It can mispronounce names, loanwords and unusual spellings.
• Output varies between runs on identical input — generation is sampled, not deterministic.
• Quality depends heavily on the reference recording. A noisy or flat sample produces a flat clone.

Automated decisions

We make none about you. No profiling, no automated decisions with legal or similarly significant effect. Abuse enforcement is decided by a person.

Questions

Reporting misuse of a voice, including your own, goes to hello@dub44.com and we act on it.